Share your thoughts, 1 month free Claude Pro on us
See more
Home
/
Benchmarks
Sparse Autoencoding on Qwen2.5-3B Layer 24 activations (held-out)
Loading...
0.591
Relative Error
Dense-SAE
0.58072
0.65011
0.7195
0.78889
Jul 2, 2026
Relative Error
Reconstruction CE
Updated 22d ago
Evaluation Results
Method
Method
Links
Relative Error
Reconstruction CE
Dense-SAE
Layer=24, m=2,048, n=1...
2026.07
0.591
0.936
Expander d=102
Layer=24, m=2,048, n=1...
2026.07
0.659
0.884
Expander d=7
Layer=24, m=2,048, n=1...
2026.07
0.693
0.875
Expander d=30
Layer=24, m=2,048, n=1...
2026.07
0.848
0.821
Feedback
Search any
task
Search any
task